Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Post-merge paper-voice gate: inspect the reader-facing draft for planner-talk and axis-label artifacts, then route fixes to the actual source. **Trigger**: post-merge voice gate, merge voice gate, transition leakage, planner talk, 合并后口吻门, 过渡句污染. **Use when**: `section-merger` has produced `output/DRAFT.md` and you want to ensure merge-injected text won't drag the draft into generator voice before polishing.
.claude/skills/willoscar-post-merge-voice-gate/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 435% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 50% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 50% | 0% |
Purpose: catch the highest-impact "automation tells" that appear after merge.
Why this exists:
section, or outline/transitions.md only when insertion was enabled.
This skill is a gate:
output/POST_MERGE_VOICE_REPORT.md).output/QUALITY_GATE.md so the workspace remains debuggable.textYou are the post-merge voice gatekeeper for a survey draft. Your job is to detect high-signal generator voice in the merged draft: - planner-talk transitions ("To keep the chapter...", "The remaining uncertainty is...") - slide/navigation narration ("Next, we move...", "We now turn...") - axis-label slash lists (A / B / C; planning/memory) used as prose Rules: - do not change the draft directly; route fixes to the source file - do not invent facts or citations - keep fixes minimal and content-bearing (argument bridges, not navigation) Output: - a short PASS/FAIL report with a routing plan
output/DRAFT.mdoutline/transitions.mdoutline/transitions.md is an optional planning artifact by default. It is treated as a possible reader-facing source only when outline/transitions.insert_h3.ok or outline/transitions.insert_h2.ok exists.
output/POST_MERGE_VOICE_REPORT.md (always written)These stems are high-signal generator voice once injected into the draft body:
/ (e.g., "retrieval / index / write policy")outline/transitions.md:transition-weaver (or hand-edit outline/transitions.md), then rerun section-merger.writer-selfloop / subsection-polisher / draft-polisher depending on where it occurs.Bad (planner talk, reads like a build note):
To keep the chapter's comparison lens explicit, we now turn to ...Better (argument bridge, content-bearing):
Once interface contracts fix what actions are executable, the next bottleneck is how agents choose among those actions under uncertainty and budget constraints.Bad (axis-label slash list):
... under mechanism/architecture/data trade-offs ...Better (natural prose):
... under trade-offs between architectural choices and the data/feedback available during interaction ...uv run python .codex/skills/post-merge-voice-gate/scripts/run.py --workspace <workspace>--workspace <dir>: workspace root--unit-id <U###>: unit id (optional; for logs)--inputs <semicolon-separated>: override inputs (rare; default: output/DRAFT.md;outline/transitions.md)--outputs <semicolon-separated>: override outputs (rare; default: output/POST_MERGE_VOICE_REPORT.md)--checkpoint <C#>: checkpoint id (optional; for logs)section-merger (recommended):uv run python .codex/skills/post-merge-voice-gate/scripts/run.py --workspace <workspace>output/POST_MERGE_VOICE_REPORT.md and appends a short record to output/QUALITY_GATE.md.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | fail→fail | 12,643 | 14,738 | +17% | 1 | 1 | 0% | 1,825 | 3,701 | +103% | 0 | 0 | — |
case-01 | pass→fail | 30,989 | 4,899 | -84% | 1 | 1 | 0% | 5,792 | 1,395 | -76% | 0 | 0 | — |
case-02 | fail→fail | 4,214 | 4,654 | +10% | 1 | 1 | 0% | 201 | 1,309 | +551% | 0 | 0 | — |
case-03 | fail→fail | 3,709 | 4,175 | +13% | 1 | 1 | 0% | 280 | 1,246 | +345% | 0 | 0 | — |
case-04 | fail→pass | 13,129 | 7,665 | -42% | 1 | 1 | 0% | 479 | 2,565 | +435% | 0 | 0 | — |
case-05 | fail→fail | 5,808 | 5,301 | -9% | 1 | 1 | 0% | 219 | 1,423 | +550% | 0 | 0 | — |
case-07 | fail→pass | 20,973 | 9,574 | -54% | 1 | 1 | 0% | 3,516 | 3,139 | -11% | 0 | 0 | — |
case-08 | fail→pass | 8,000 | 4,870 | -39% | 1 | 1 | 0% | 1,305 | 2,089 | +60% | 0 | 0 | — |
case-09 | fail→pass | 7,855 | 4,312 | -45% | 1 | 1 | 0% | 1,271 | 1,901 | +50% | 0 | 0 | — |
case-10 | fail→pass | 8,321 | 4,346 | -48% | 1 | 1 | 0% | 1,272 | 1,914 | +50% | 0 | 0 | — |
case-11 | fail→pass | 11,542 | 3,437 | -70% | 1 | 1 | 0% | 2,273 | 1,687 | -26% | 0 | 0 | — |
case-12 | fail→fail | 8,270 | 2,934 | -65% | 1 | 1 | 0% | 1,252 | 1,579 | +26% | 0 | 0 | — |
case-13 | fail→pass | 18,828 | 1,591 | -92% | 1 | 1 | 0% | 991 | 1,282 | +29% | 0 | 0 | — |
case-14 | pass→pass | 6,385 | 5,109 | -20% | 1 | 1 | 0% | 981 | 1,772 | +81% | 0 | 0 | — |
case-15 | pass→pass | 5,518 | 2,348 | -57% | 1 | 1 | 0% | 789 | 1,429 | +81% | 0 | 0 | — |
case-16 | fail→pass | 8,108 | 2,434 | -70% | 1 | 1 | 0% | 1,437 | 1,468 | +2% | 0 | 0 | — |
case-17 | fail→pass | 22,076 | 2,293 | -90% | 1 | 1 | 0% | 1,097 | 1,395 | +27% | 0 | 0 | — |
case-18 | pass→pass | 8,273 | 3,827 | -54% | 1 | 1 | 0% | 1,520 | 1,706 | +12% | 0 | 0 | — |
case-19 | pass→pass | 10,814 | 5,578 | -48% | 1 | 1 | 0% | 1,585 | 1,943 | +23% | 0 | 0 | — |
case-20 | fail→pass | 10,687 | 3,645 | -66% | 1 | 1 | 0% | 1,767 | 1,766 | -0% | 0 | 0 | — |
case-21 | pass→pass | 8,702 | 2,015 | -77% | 1 | 1 | 0% | 1,372 | 1,384 | +1% | 0 | 0 | — |
case-22 | fail→pass | 7,760 | 2,061 | -73% | 1 | 1 | 0% | 1,289 | 1,400 | +9% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 16 counted toward the lift figure. The other 6 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 16 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.